Landscape of Transcriptional Deregulation in Lung Cancer Shu Zhang1,2,3,4, Mingfa Li1, Hongbin Ji2,3,4,5* and Zhaoyuan Fang2,3,4,6*

Zhang et al. BMC Genomics (2018) 19:435 https://doi.org/10.1186/s12864-018-4828-1

RESEARCHARTICLE Open Access Landscape of transcriptional deregulation in lung cancer Shu Zhang1,2,3,4, Mingfa Li1, Hongbin Ji2,3,4,5* and Zhaoyuan Fang2,3,4,6*

Abstract Background: Lung cancer is a very heterogeneous disease that can be pathologically classified into different subtypes including small-cell lung carcinoma (SCLC), lung adenocarcinoma (LUAD), lung squamous cell carcinoma (LUSC) and large-cell carcinoma (LCC). Although much progress has been made towards the oncogenic mechanism of each subtype, transcriptional circuits mediating the upstream signaling pathways and downstream functional consequences remain to be systematically studied. Results: Here we trained a one-class support vector machine (OC-SVM) model to establish a general transcription factor (TF) regulatory network containing 325 TFs and 18724 target genes. We then applied this network to lung cancer subtypes and identified those deregulated TFs and downstream targets. We found that the TP63/SOX2/ DMRT3 module was specific to LUSC, corresponding to squamous epithelial differentiation and/or survival. Moreover, the LEF1/MSC module was specifically activated in LUAD and likely to confer epithelial-to-mesenchymal transition, known important for cancer malignant progression and metastasis. The proneural factor, ASCL1, was specifically up-regulated in SCLC which is known to have a neuroendocrine phenotype. Also, ID2 was differentially regulated between SCLC and LUSC, with its up-regulation in SCLC linking to energy supply for fast mitosis and its down-regulation in LUSC linking to the attenuation of immune response. We further described the landscape of TF regulation among the three major subtypes of lung cancer, highlighting their functional commonalities and specificities. Conclusions: Our approach uncovered the landscape of transcriptional deregulation in lung cancer, and provided a useful resource of TF regulatory network for future studies. Keywords: Lung cancer, Transcription factors, Support-vector machines, Transcription regulatory network

Background systematically uncover the molecular regulatory network Lung cancer is the leading cause of cancer-related deaths in contributing to lung cancer malignant progression. worldwide. Pathologically, lung cancers can be classified Transcription factors (TFs), known to be evolutionarily as small-cell lung carcinoma (SCLC) and non-small-cell conserved in orchestrating transcriptional gene regulation lung carcinoma (NSCLC), and the latter can be further networks, are the key players in contribution to a broad divided into lung adenocarcinoma (LUAD), lung squa- range of critical cellular physiological and pathological mous cell carcinoma (LUSC), and others such as processes, from normal development and physiological large-cell carcinoma (LCC). Among these lung cancer processes to diseases such as cancer [4–7]. Notably, mas- subtypes, LUAD, LUSC and SCLC are most prevalent, ter TFs bind to the corresponding promoter regions via accounting for about 40%, 25-30% and 10-15% respect- recognizing specific short sequence patterns (‘motifs’), and ively (https://www.cancer.org). Previous mechanistic regulate transcriptional expression of a series of target studies have greatly advanced our knowledge about genes, which thus control cell growth, proliferation and how lung cancer initiates, progresses and responds to differentiation. For instance, TFs such as PPARγ and C/ drug treatments [1–3]. However, it remains interesting to EBPα are key regulators of adipogenic differentiation [8]. Overexpression of TFs including OCT4, SOX2, KLF4 and MYC can reprogram fibroblasts to pluripotent stem cells * Correspondence: [email protected]; [email protected] [9, 10]. Nanog, another TF which is transcriptionally regu- 2State Key Laboratory of Cell Biology, Shanghai, China Full list of author information is available at the end of the article lated by OCT4 and SOX2, is also important for the

© The Author(s). 2018 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated. Zhang et al. BMC Genomics (2018) 19:435 Page 2 of 13

maintenance of pluripotency [11]. Furthermore, TFs transformation and quantile normalization with voom [26], are the major driving forces of transdifferentiation and Pearson Correlation Coefficient (PCC) was computed for transition among different cell types [12]. Such TF regula- each pair of genes. Protein-protein interactions were down- tory programs also exist in cancer. For example, the loaded from the integrated EBI IntAct molecular inter- epithelial-to-mesenchymal transition (EMT) process, action database [27]. For each candidate gene, its PCCs mediated by key TFs such as SNAILs and bHLHs, is with the TF and TF-interacting proteins (‘neighbors’)were known to promote cancer malignant progression and computed, and the latter PCCs were summarized into three metastasis [13, 14]. The reprogramming factor, SOX2, quantiles (25% as Q1, 50% as M, 75% as Q3). The candidate has also been identified as a lineage-survival oncogene gene’s PCCs with the background genes were also calcu- in LUSC [15]. SOX2 and TP63 (the other known LUSC lated and summarized into these three quantiles. lineage TF) are both frequently amplified and crucial for LUSC development [15–17]. Recently, we have also OC-SVM model training and evaluation shown that, TP63 mediates the transdifferentiation from One-class support vector machine (OC-SVM) is a spe- LUAD to LUSC [18]. cial type of SVM model suitable for solving problems To systematically understand how transcription factors where high-quality training data is available for only one contribute to the malignant progression of lung cancer, class, and it has been widely used in single-class learning we employed a machine learning approach to build a and outlier detection [28, 29]. Here we used curated transcriptional regulatory network, based on curated regu- TF-target relations from the TRRUST database as the latory relations, motif distributions, protein-protein interac- positive training set [30], with synthetic negatives to tions (PPIs) and gene co-expression. With the application evaluate the model performance. The negative set was of this network in LUSC, LUAD and SCLC, we identified built with 1000 20kb random sequences scanned with those core TFs specific for each lung cancer subtype. We FIMO using the same setting. The correlation coefficient further described the landscape of TF deregulation in these data for synthetic genes were randomly chosen from real three major lung cancer subtypes. gene correlation coefficients. A random subset of 50,000 TF-target pairs were used for evaluation. The OC-SVM Methods model was trained using the libSVM R wrapper in the Lung cancer data sources and preprocessing e1071 package. With the radial basis kernel and a series The RNA-Seq FPKM and copy number data for TCGA of ‘nu’ (ranging between 1^-4 and 0.9) and ‘gamma’ LUAD and LUSC were downloaded from the UCSC Xena (2^-5, 2^-8, 2^-11), the performance of models were hub (http://xena.ucsc.edu/). The SCLC gene expression data assessed in terms of sensitivity and false positive rate were obtained from the paper-accompanied data [19]. Other (FPR) with 10-fold cross-validation. To achieve a high LUAD and LUSC data outside of TCGA were downloaded specificity that is essential for large-scale predictions from the NCBI GEO with accession number GSE81089. To where the candidate relations are huge (over 17,000,000), be concise, we refer to these LUAD and LUSC datasets out- we controlled the final model (nu=0.5, gamma=2^-5) at a side of TCGA as ‘LUAD2’ and ‘LUSC2’. For FPKM data, a relatively low FPR (0.002), sacrificing some sensitivity log-transformation was applied before downstream analyses (50%). This predicted 2,432,769 relationships between TFs of co-expression and differential expression. and protein-coding target genes, and ~5000 of them were likely to be false positives. Promoter sequences and motif analyses We obtained genomic sequences (UCSC hg19) from 10kb Identification of core TFs in lung cancer upstream to 10kb downstream of TSS for each Ensembl To ensure specificity on the lung cancer dataset, we fil- gene. Non-redundant TF motifs were from the JASPAR tered the predicted targets for individual TFs by enfor- database [20] and converted to MEME format. Additional cing two sequential steps: (i) the target gene must have motifs (NKX2-1 and ASCL1) were trained from the re- conditional co-expression with the TF (PCC>=0.5); (ii) ported TF binding peaks [21, 22], with the MEME-ChIP the target gene must have inter-correlations with at least pipeline [23]. Scanning of motifs along promoter sequences 1/6 of the other target genes (PCC>=0.5). Thus we en- was performed with FIMO (default p value threshold, 1e-4) sured both the TF-target correlations and the overall [24]. FIMO matches on each strand were categorized by inter-correlations among the targets. We next deter- upstream 10kb, 2kb, 500b and downstream 10kb, 2kb, mined the differential regulation of TF and targets in 500b, respectively. cancer versus normal tissue. A 2-fold expression change threshold (i.e. log2fc=1) and paired Student’s T test were Gene co-expression and network neighborhood analyses used to determine up- and down-regulated genes. The We downloaded the comprehensive tissue profiling data Benjamini-Hochberg method was used to control the from the GTEx project (version v6p) [25]. After logarithmic overall false discovery rates (FDR=0.1). All datasets were Zhang et al. BMC Genomics (2018) 19:435 Page 3 of 13

analyzed with these same threshold settings. For the TFs, we only required them to be weakly differentially expressed in cancer versus normal (log2fc>=0.3 and p<=0.05), as we noticed some TFs may not be very strongly deregulated at the mRNA level. Then, for each TF, we counted the number of its target genes that were up- and down-regulated in cancers (‘n_up’ and ‘n_down,re-’ spectively), and classified the TF-targets group as ‘up’ if the TF was overexpressed and n_up/n_down>=10 (vice versa).

Gene Ontology analysis Gene Ontology (GO) annotations for human were obtained from the org.Hs.eg.db package (Bioconductor). The GO hierarchy was downloaded from the GO official website (http://geneontology.org) and we focused on the ‘biological processes’ category, which are more relevant to functional enrichment analysis. Fisher’s exact test was used to assess the enrichment for each GO term, and those significant terms (p<0.05 and OR>2) were further filtered according to the GO hierarchy with a priority given to more specific terms.

Results An OC-SVM model for predicting transcriptional Fig. 1 Prediction of TF targets with OC-SVM. TF binding motifs were scanned along promoter regions (-10kb~+10kb around TSS) for regulatory network annotated genes. Co-expression between TF and candidate To unravel the TF regulatory network in the major lung targets, as well as between the TF PPI neighborhood and cancer subtypes, we designed a two-step strategy: first candidate targets, were analyzed. An OC-SVM model was trained build an overall TF regulatory network, and then com- with curated TF-target knowledge, and synthetic negatives were bine dataset information to identify dataset-specific TFs used for evaluating its performance and regulation. Over the years, experimentally validated TF-target relationships have accumulated and become a with GADD45A, its interacting proteins indeed showed valuable resource for learning general principles that would strong positive co-expression with GADD45A. Therefore guide further discoveries of novel regulation [30–32]. For we integrated the neighborhood co-expression with target such experimental knowledge, the positive training datasets genes into the OC-SVM model. are of high quality whereas the negative datasets are mostly To assess the performance of the OC-SVM model, we unavailable. To build a global TF regulatory network artificially synthesized some negative sets based on the based on the resource available, we took advantage of following principles: 1) the synthetic genes’ promoter an OC-SVM framework that has been widely used in regions are randomly generated and then summarized the single-class prediction field [33]. for individual TF-binding motifs; 2) the co-expression We collected and extracted the following information between synthetic genes and other genes including TFs for establishing TF-target relationships: the presence and and TF neighbors were randomly extracted from real distribution of TF binding motifs along the promoter co-expression data using a randomized gene label. Model regions, the co-expression between a TF and its target genes, performance was evaluated with 10-fold cross-validation. as well as the co-expression of a TF’s interacting proteins At a sensitivity level of 75%, the true positive rates are gen- (‘neighborhood’) with its target genes (Fig. 1, Methods). erally above 90% (Fig. 3a). We realized that minimizing the From the distribution of Pearson correlation coefficients FPR was critical for our tasks, since the number of possible (PCCs), there was much stronger positive co-expression regulatory relationships are rather huge: e.g. for 300 TFs than the background (Fig. 2a, b), implicating the rationality and 20,000 genes, there would be 6 million possible rela- of co-expression-based TF-target prediction. In addition, tions. Therefore we had to minimize FPR as long as the the TF-interacting proteins displayed a positive but weaker sensitivity was acceptable. To further guarantee the appro- co-expression with target genes. An interesting example priate choice of model parameters, we evaluated different was JUND, which regulated downstream target gene parameter combinations (nu=0.3, 0.5, 0.7; log2gamma=-5, GADD45A (Fig. 2c-d, Additional file 1: Figure S1). Al- -8, -11) for TF network training, with a real dataset (TCGA though JUND itself did not show clear co-expression LUSC) and two known core LUSC TFs (TP63 and SOX2) Zhang et al. BMC Genomics (2018) 19:435 Page 4 of 13

Fig. 2 Co-expression analyses for TF, TF neighborhood and known target genes. a, b Distribution of PCCs between TFs and target genes, between TF neighborhoods and target genes, and among all genes as the background. c JUND and its neighborhood network. Nodes were colored according to co-expression with JUND’s known target GADD45A. d Co-expression distribution between JUND’s neighborhood and GADD45A serving as positive controls. Each combination successfully predicted targets; (ii) co-expression among the predicted recalled both TFs, indicating that core TFs might be identi- targets; (iii) differential regulation between cancer and nor- fied even with a less sensitive model (Additional file 2: mal tissue: the TF itself should at least be weakly deregu- Table S3). Nonetheless, the number of targets predicted for lated and its targets should be distributed in the same each TF decreased with lower model sensitivities, empha- directionastheTF,withanenrichmentof10foldversus sizing that a higher model sensitivity might be more power- the opposite direction (Methods). ful to detect core TFs (Additional file 2: Table S3). Based on In order to evaluate the effect of differential criteria on the cross-validation and real dataset evaluations above, we TF identification, various combinations of log2fc and chose an appropriate parameter combination (nu=0.5 and FDR q value thresholds were tried on the TCGA LUSC log2gamma=-5) to balance our specific requirements of dataset. Although the numbers of up- and down-regulated sensitivity (~50%) and FPR (~0.2%). This resulted in a pre- genes fluctuated greatly, the TFs identified were quite dicted network of 325 TFs and 18724 protein-coding target stable, indicating the robustness of the methodology genes (Fig. 3b). The numbers of target genes for TFs (Additional file 2: Table S4). Therefore, the same differen- are 7332 in median (ranging from 338 to 15929), and tial threshold (|log2fc|>=1 and q<=0.1) was applied to all the numbers of regulatory TFs for genes are 139 in me- datasets. dian (ranging from 0 to 244), indicating the network We applied the above analyses and requirements to the was quite general and should be narrowed down for following lung cancer datasets (Methods), and identified identification of condition-specific regulation. dataset-specific regulatory TFs: TCGA LUAD (referred to as ‘LUAD’), TCGA LUSC (referred to as ‘LUSC’), SCLC Identification of dataset-specific differential dataset (referred to as ‘SCLC’), independent LUAD and transcriptional regulation LUSC dataset (referred to as ‘LUAD2’ and ‘LUSC2’ re- To identify condition-specific regulation, we enforced three spectively) (Additional file 2: Table S1). We also clustered requirements (Methods): (i) co-expression between TF and the up- and down-regulated TFs according to their targets Zhang et al. BMC Genomics (2018) 19:435 Page 5 of 13

Fig. 3 Training and prediction of the OC-SVM model. a ROC curves for model evaluation with 10-fold cross validation. The positive sets were curated known TF-target regulatory relationships, whereas the negative sets were artificially synthesized (See Methods). ROC curves for three values of log2 gamma parameter were shown: -11, -8, -5. b Predictions of OC-SVM. Left, distribution of TFs by the number of predicted targets. Right, distribution of genes by the number of TFs predicted to target them overlapping to identify potential co-regulated TFs (Fisher’s [15–17, 34–36]. Moreover, our analyses indicated that exact test, p <0.05). DMRT3 was associated with TP63 and SOX2 in the same module (Fig. 4b-d). The functional implication The TP63/SOX2/DMRT3 circuit as a hallmark of lung of DMRT3 in LUSC was not well known, though two squamous carcinomas earlier studies found that DMRT3 could be lost We identified 26 up-regulated TFs in LUSC, 21 of which through copy number alteration mechanisms in LUSC were also identified in the LUSC2 dataset independently, [37, 38]. To reconcile this seeming discrepancy, we suggesting a good agreement between different datasets exploited inter-correlations among DMRT3 copy (Fig. 4a, Additional file 3: Figure S2A, Additional file 2: number, DMRT3 expression, and TP63/SOX2 expres- Table S1). We then merged these two sets of up-regulated sion through an integrative analyses of the TCGA data. TFs and only retained those with shared target genes. A We found that the copy number status of DMRT3 was further clustering of these TFs showed some of them were heterogeneous in LUSC, with tumors not bearing well clustered into TF modules (Fig. 4b,Additionalfile3: DMRT3 deletions having significantly higher DMRT3 Figure S2B). expression, as well as significantly increased TP63/ Among these, TP63 and SOX2 were well-known SOX2 expression (Additional file 3:FigureS2C-E). LUSC-specific oncogenic TFs that were important in These indicated that DMRT3 might have dual func- squamous epithelial differentiation and/or survival tions correlated with the heterogeneity of LUSC, with Zhang et al. BMC Genomics (2018) 19:435 Page 6 of 13

c d

Fig. 4 Transcriptional hallmarks for LUSC. a Consistency of up-regulated TFs identified in the LUSC and LUSC2 datasets. b Clustering of up- regulated TFs shared in the two LUSC datasets. TFs with 10 or fewer targets shared between the two datasets have been filtered out before clustering. Cluster membership was determined using Fisher’s exact test (p<0.05). c, d Expression patterns of the TP63/SOX2/DMRT3 module and their commonly regulated genes in LUSC (c) and LUSC2 (d) datasets. e Functional enrichment of co-regulated genes by TP63/SOX2/DMRT3 (left). A hypothetical regulatory model was proposed (right)

its higher expression mainly restricted to samples overex- LUSC development and squamous phenotype formation pressing TP63/SOX2. In addition, both SOX2 and DMRT3 (Fig. 4e, right). Interestingly, a more recent study identified targeted the TP63 promoter (Additional file 3:FigureS2F), DMRT3 as an important regulator of neuronal differenti- and these three factors altogether co-regulated a common ation programs involved in locomotor network develop- subset of genes involved in epithelial cell differentiation ment [39]. Future experimental studies are worth to fully (Fig. 4e,left).Therefore,wehypothesizethatDMRT3may characterize the implication of DMRT3 with SOX2/TP63 participate in the TP63/SOX2 circuit for regulating squa- in augmenting LUSC epithelial survival. mous cell differentiation and/or survival, and that these Furthermore, a comparison with the other two lung can- three factors may co-regulate genes functioning in human cer subtypes revealed that, the TP63/SOX2/DMRT3 circuit Zhang et al. BMC Genomics (2018) 19:435 Page 7 of 13

was among the TFs up-regulated in a LUSC-specific man- MSC and LEF1 might functionally converge at EMT. In ner (Fig. 7c), consistent with known properties of squamous LUAD, MSC and LEF1 clustered together to regulate a lineage survival TFs. shared set of target genes (Fig. 5b). Furthermore, analyses of these genes co-regulated by MSC and LEF1 revealed Functional regulation transcriptionally encoded in lung significant enrichment of terms such as extracellular adenocarcinomas matrix (ECM) organization and cell-ECM interactions, We next analyzed the TF modules that were up-regulated which were related to EMT (Fig. 5c, d). Together, our data in LUAD (Fig. 5). The two independent datasets again showed that two LUAD-specific TFs, MSC and LEF1, show good agreement, although not as good as that in might synergize in promotion of lung cancer malignant LUSC datasets (Fig. 5a). To reduce batch effects, we re- progression through EMT process. stricted our analyses to the LUAD dataset. Several LUAD Surprisingly, NKX2-1, a TF amplified in about 12% of TFs were commonly shared with LUSC, such as E2F7, LUAD [43], turned out to be a down-regulated regulator in E2F8, MYBL2, TFAP2A, TFAP4 and OTX1 (Fig. 4b, 5b, the TCGA LUAD dataset, and not identified in the LUAD2 Additional file 2: Table S1). Other TFs such as LEF1 dataset (Additional file 4:FigureS3B,Additionalfile5: (Lymphoid Enhancer-binding Factor 1) and MSC (Muscu- Figure S4, Additional file 2: Table S1). Several observa- lin, also Activated B-Cell Factor 1) were specific to LUAD tions might help explain this unexpected result. First, and not present in LUSC or SCLC (Fig. 7c, Additional NKX2-1 was amplified in only a limited subset of file 2: Table S1). LEF1 is in the Wnt signaling pathway LUAD tumors (Additional file 4: Figure S3C) [43]. Second, and known to regulate the EMT process. It has been NKX2-1 expression showed a stage-dependent manner, found to be activated in multiple cancer types ranging with up-regulation in stage I and gradual down-regulation from leukemia to solid tumors including LUAD [40]. from stage II to IV (Additional file 4: Figure S3D), in con- Consistent with its function in EMT, LEF1 drives me- sistent with previous publication [44]. Third, it has been tastasis of primary LUAD to brain and bone [41]. The proposed that NKX2-1 plays dual roles in LUAD, both other factor, MSC, is less studied in lung cancer. Nonethe- oncogenic and anti-oncogenic (also anti-metastatic) in less, its overexpression has been implicated in disruption LUAD [45, 46]. Taken together, NKX2-1 may have of normal B cell differentiation program and Hodgkin stage-specific function in LUAD and tends to be lymphoma development [42]. These data suggest that down-regulated as LUAD become advanced.

Fig. 5 Transcriptional deregulation in LUAD. a Consistency of up-regulated TFs identified in the LUAD and LUAD2 datasets. b Clustering of up- regulated TFs identified in the TCGA LUAD dataset. Cluster membership was determined using Fisher’s exact test (p<0.05). c Expression pattern of the LEF1/MSC module and their common targets in TCGA LUAD dataset. d Functional enrichment of genes co-regulated by LEF1/MSC Zhang et al. BMC Genomics (2018) 19:435 Page 8 of 13

Regulatory patterns specific to small-cell lung carcinomas organization, mitochondrion protein translations and Traditionally, LUAD and LUSC are categorized in the ATP synthesis (Fig. 6c), and its up-regulation probably NSCLC group, as SCLC is distinct in its cell size, shape and assisted SCLC cells in gaining sufficient energy to support cell mitosis rate. In SCLC, we found those uniquely fast mitosis and proliferation. However, in LUSC, ID2 con- up-regulated TFs such as ASCL1, CENPB, HSF2, ZNF143 ditionally regulated another set of genes involved in positive and down-regulated TFs such as STAT3, REST, NFKB1, regulation of immune response, leukocyte cell activation different from those in LUAD and LUSC (Fig. 6a-b,Fig.7c, and immune signaling (Fig. 6d), and down-regulation of Additional file 2: Table S1). Among these, the bHLH family ID2 and its target genes help LUSC cells to escape immune TF ASCL1, a well-known neuronal differentiation regulator, surveillance. This indicated that different types of cancer is required by neuroendocrine tumors including SCLC cells may deregulate the same TF differently, in support of [47–49]. ASCL1 target genes showed an involvement in cancer-specific need in malignant progression. regulation of neurotransmitter levels and presynaptic process related to synaptic transmission (Additional file 2: The transcriptional regulatory landscape of lung cancer Table S2). Moreover, the target genes of ASCL1 were sig- subtypes nificantly shared by FOXA2, whose target genes were also We have unraveled the key TFs as well as their targets in enriched for neural-related functions including neuronal each of the three major subtypes of lung cancer (Fig. 7c, generation and cell migration (Additional file 2: Table S2). Additional file 5: Figure S4, Additional file 2:TableS1). These again emphasized the unique neuroendocrine fea- Notably, there were some deregulated TFs shared by all tures of SCLC, in contrast to LUAD and LUSC. three subtypes. For example, two TFs, E2F1 and TCF3, Interestingly, some TFs showed opposite expression were up-regulated in all three subtypes (Fig. 7a, c). These changes in comparison with LUAD and/or LUSC. For two factors both regulated target genes mainly involved in example, ID2, FOXA2 and ID4 were up-regulated in cell cycle and/or cell division processes (Additional file 2: SCLC but down-regulated in LUAD and/or LUSC. Table S2). We found that E2F1 regulated genes enriched in Similarly, TP63 and RARG were down-regulated in ‘cell division’ across all three subtypes, with three target SCLC but up-regulated in LUSC (Fig. 7c). We next ex- genes in the GO term commonly regulated in lung cancers: plored the potentially opposite roles of ID2 in SCLC CCNF (cyclin F), NCAPH (Non-SMC Condensin I Com- and LUSC. In SCLC, ID2 regulates mitochondrion plex Subunit H), SPAG5 (Sperm Associated Antigen 5).

Fig. 6 Transcriptional deregulation in SCLC. a-b Clustering of up-regulated (a) and down-regulated (b) TFs, respectively. Cluster membership was determined using Fisher’s exact test (p<0.05). c Functional enrichment of ID2 target genes in SCLC. d Functional enrichment of ID2 target genes in LUSC Zhang et al. BMC Genomics (2018) 19:435 Page 9 of 13

Fig. 7 Landscape of transcriptional deregulation in lung cancer. a Comparison of up-regulated TFs in LUAD, LUSC and SCLC datasets. b Comparison of down-regulated TFs in LUAD, LUSC and SCLC datasets. c The global patterns of TF deregulation across the five datasets: LUAD, LUAD2, LUSC, LUSC2 and SCLC. Colors reflect the log2 scaled number of a TF’s targets, with up-regulated TFs in red and down-regulated in blue. Selected branches of TFs that were common (orange for NSCLC-common, yellow for all-common) or subtype-specific (blue) are highlighted (bottom)

Moreover, five TFs were found to be down-regulated in E2F1, TCF3) and 21 down-regulated factors (ID4, RXRG, all three subtypes: FOS, GATA2, SOX17, TBX5, TCF21 JDP2, MITF, SPI1, NFIX, NR2F1, ZEB1, ZNF423, ERG, (Fig. 7b, c). They regulate various functions ranging TFEC, ETS1, HOXA5, PKNOX2, TCF21, FLI1, SOX17, from ‘inflammatory response’ to ‘positive regulation of TBX5, IRF8, FOS, GATA2). The up-regulated TFs mainly apoptotic process’. Some TFs shared the same target regulated cell proliferation (‘mitotic nuclear division,’ ‘cell genes across the different subtypes, e.g., FLI1 probably tar- division,’ ‘G1/S transition of mitotic cell cycle’ and ‘DNA re- gets CCRL2 (Chemokine/C-C Motif Receptor-Like 2), an pair’), and the down-regulated TFs mainly regulated cell essential regulator of leukocyte recruitment in the lung differentiation (‘mesenchymal cell differentiation,’ ‘lung de- [50], in all three subtypes. velopment,’‘embryonic morphogenesis,’‘pattern specification We also found dramatic difference of regulation patterns process’), cell proliferation (‘negative regulation of cell pro- among the subtypes. The two NSCLC isoforms (LUAD and liferation’) and immune responses (‘inflammatory response,’ LUSC) shared more TFs than with SCLC (Fig. 7a, b). LUAD ‘T cell proliferation,’ ‘T cell aggregation’) (Additional file 2: and LUSC shared 5 up-regulated (TFAP4, OTX1, E2F8, TableS2).SCLCspecificallyup-regulatedaseriesofTFs Zhang et al. BMC Genomics (2018) 19:435 Page 10 of 13

(ASCL1,FOXA2,ID2,ID4,THAP1,ATF4,CENPB, enhance the specificity of target gene predictions, it is ZNF143, HSF2, ESRRA, TBP, INSM1, PKNOX1) that useful to incorporate expression relevance between TF functioned in neural functions (‘regulation of neuro- and targets [62, 63]. However, as TFs may often be regu- transmitter levels’, ‘presynaptic process’, ‘generation of lated by post-translational modifications, translocations, neurons’, ‘neuron development’, ‘neurological system as well as protein-protein interactions, its expression level process’), mitochondrial activities (‘mitochondrion could not fully represent the regulatory activity. To remedy organization’, ‘mitochondrial translational elongation’), this, we used a network-based approach to incorporate protein synthesis (‘translation’, ‘rRNA processing’), me- expression relevance dispersed in the TF neighborhood. tabolism (‘purine ribonucleoside metabolic process’) Through the integration of PWM matching, expression and cell proliferation (‘mitotic cell cycle process’, ‘cell correlations, and neighborhood relevance, an OC-SVM division’). Those down-regulated TFs in SCLC (JUNB, model was trained and evaluated for the performance in NFKB1, VENTX, CREB3L1, REST, RARB, FOXO1, predicting known targets, which allowed us to control the EGR1, TP63, ZBTB7A, STAT3, MEOX1, FOSL2, RARG, false discovery rate to 0.002. GATA5, RXRA, NPAS2, LEF1, BCL6, TCF12) were func- Another major motivation of this work is to present tionally linked to cell differentiation (‘positive regulation of the landscape of transcriptional deregulation of lung cell differentiation,’ ‘epithelial cell differentiation’)andim- cancer including three major subtypes LUAD, LUSC and mune responses (‘inflammatory response,’ ‘T cell aggrega- SCLC. We reveal those common regulatory relationships tion,’ ‘positive regulation of cytokine production, ‘leukocyte as well as subtype-specific regulatory relationships. We migration’)(Additionalfile2:TableS2).Thesefindingsin- have distinguished up- and down-regulation of TF circuits dicatedthatNSCLCandSCLChijackeddifferentmolecular in each subtype, and predicted a number of subtype-specific machineries to promote malignant progression. Nonethe- TF modules (e.g. TP63/SOX2/DMRT3, LEF1/MSC, ASCL1 less, SCLC had more specific TF circuits to increase mito- and ID2). Moreover, we have interpreted each module to chondrial activities and protein synthesis, which probably functionally explain that different mechanisms are hijacked provided high levels of cellular energy in support of fast by different cancer cells to achieve corresponding malignant mitosis [51]. progression. Notably, many of these functional outputs are A notable difference of TF circuits was even detected highly correlated, such as cell proliferation, dedifferentiation between LUAD and LUSC, two major subtypes of NSCLC. and immune suppression. Nonetheless, different subtypes of LUAD specifically up-regulated several TFs (LEF1, E2F3, lung cancer also harbor unique TF machinery in contribu- HLTF, FOXP3), whereas LUSC preferentially up-regulated tion to tumor growth. For example, in SCLC, many unique other TFs (SOX2, TP63, DMRT3, PITX1, E2F7, TFAP2A, TF circuits are related to mitosis, protein synthesis, mito- MYBL2, HOXA10, HOXC13, RARG, TFAP2C, POU6F2, chondrial activities and energetic metabolism, which are HOXD13, PAX9, TP73, E2F2). Besides the common certainly important for promoting fast cell division. The function enriched for these two up-regulated sets of epithelial differentiation programs are also dramatically ele- LUAD- and LUSC-specific TFs (‘mitotic nuclear division,’ vated in LUSC, which are known important for squamous ‘cell proliferation’), there were unique functions enriched cell lineage survival from studies of cell lines and mouse for LUSC (‘epithelial cell differentiation,’ ‘epidermis devel- models. opment,’ ‘skin development’) (Additional file 2: Table S2), There are also some limitations of this study. We have and the TP63/SOX2/DMRT3 cluster was closely related not necessarily required a TF itself to be co-expressed to this squamous differentiation program. with its target genes when training the general regulatory network. However, during the dataset analyses, we Discussion stillrequiretheTFtohave at least weak expression Transcriptional regulation serves as the fundamental changes (through using less stringent thresholds), as we regulatory program in orchestrating normal develop- want to focus on those TFs that can be regulated at ment and disease progression. To unravel the transcrip- expression level, which is also common for many TFs tional target genes of TFs, both experimental techniques important in the regulation of differentiation. Nonethe- (e.g. SELEX, ChIP-on-chip, ChIP-seq) and computa- less, this may miss some TFs that are transiently regulated tional methods have been successfully developed. Trad- without long-term changes in expression. In addition, we itionally, TF binding preferences can be characterized as restrict our analyses to activating TFs that up-regulate tar- position-weight matrices (PWMs), which are then used get genes, but the number of TFs that are repressive is to scan the promoter regions for potential hits. Although also nonnegligible. Future work will be needed to integrate PWM-based methods and extensions have been widely them into a more flexible model. Moreover, the SCLC followed and deeply exploited [52–59], sequence-based dataset that we used lacks normal controls, and so we methods per se are not sufficient to account for the full used the adjacent normal samples in the LUAD and LUSC TF-DNA interaction specificities in vivo [60, 61]. To datasets to compare with SCLC. Although those adjacent Zhang et al. BMC Genomics (2018) 19:435 Page 11 of 13

normal tissues from LUAD and LUSC are quite similar datasets. (B) Clustering of down-regulated TFs shared in the two LUSC (Additional file 6: Figure S5), we cannot rule out the possi- datasets. Cluster membership was determined using Fisher’s exact test bility that those from SCLC might be different. (p<0.05). (C) DMRT3 expression grouped by DMRT3 copy number status (deletion vs. non-deletion) (Wilcoxon signed-rank test). (D) DMRT3 loss The complete landscape of complex deregulation in status in relation to TP63 expression (Wilcoxon signed-rank test). (E) various lung cancer subtypes still contains many gaps DMRT3 loss status in relation to SOX2 expression (Wilcoxon signed-rank and missing parts. This work provides an initial compre- test). (F) SOX2 (red) and DMRT3 (blue) binding motifs on the TP63 promoter (-10kb to +10kb of TSS). Genomic coordinates are according to the hg19 hensive study to unravel the overall patterns with an em- assembly. (PDF 92 kb) phasis on those important circuits in lung cancer. Future Additional file 4: Figure S3. Down-regulation of TFs in LUAD. (A) studies from both computational and experimental ap- Consistency of down-regulated TFs identified in the LUAD and LUAD2 proaches would be necessary to decode and validate the datasets. (B) Clustering of down-regulated TFs identified in the TCGA LUAD dataset. Cluster membership was determined using Fisher’s exact transcriptional networks in various lung cancer subtypes, test (p<0.05). (C) NKX2-1 copy number distribution in TCGA-LUAD dataset. including those not covered here, such as LCC. (D) NKX2-1 expression in normal lung and LUAD categorized by tumor stage (I to IV). (PDF 94 kb) Additional file 5: Figure S4. The complete version of Fig. 7c, showing Conclusions the global TF deregulation patterns across the five datasets: LUAD, We have systematically studied the core transcriptional LUAD2, LUSC, LUSC2 and SCLC. Colors reflected the log2 scaled number deregulation in three well-characterized lung cancer of a TF’s targets, with up-regulated TFs in red and down-regulated in subtypes (LUAD, LUSC and SCLC), and identified a blue. (PDF 22 kb) Additional file 6: Figure S5. Consistency among the normal lung number of common (e.g. proliferation-related E2F1 tissues from the four datasets: TCGA-LUAD, TCGA-LUSC, LUAD2 and and TCF3) as well as subtype-specific TF circuits (e.g. LUSC2. The PC1 and PC2 axes from Principal Component Analysis (PCA) the epithelial-development-related TP63/SOX2/DMRT3 together explained 91.8% of total variance. A good consistency of these normal lung tissues justified the assumption that they could be pooled module in LUSC, the EMT-related LEF1/MSC module in together for comparison with SCLC cancer samples. (PDF 42 kb) LUAD, and the neural differentiation regulator ASCL1 in SCLC). Moreover, ID2 targets two different sets of genes Abbreviations with one involved in mitochondrial activities in SCLC CCNF: Cyclin F; CCRL2: Chemokine/C-C Motif Receptor-Like 2; and the other involved in immune response in LUSC, ECM: Extracellular matrix; EMT: Epithelial-to-mesenchymal transition; highlighting the importance of the same TF differen- FDR: False discovery rate; FPR: False positive rate; GO: Gene Ontology; LCC: Large-cell carcinoma; LEF1: Lymphoid Enhancer-binding Factor 1; tially regulated in different cancer subtypes. Nonethe- LUAD: Lung adenocarcinoma; LUSC: Lung squamous cell carcinoma; less, different TFs are also employed by NSCLC and MSC: Musculin; NCAPH: Non-SMC Condensin I Complex Subunit H; SCLC to achieve similar functional consequences to NSCLC: Non-small-cell lung carcinoma; OC-SVM: One-class support vector machine; PCC: Pearson Correlation Coefficient; PPI: Protein-protein support tumor progression. interaction; PWM: Position-weight matrix; SCLC: Small-cell lung carcinoma; SPAG5: Sperm Associated Antigen 5; TF: Transcription factor Additional files Acknowledgements The authors would like to thank Yu Zhang for help in an initial draft of this Additional file 1: Figure S1. Co-expression between JUND or its work and Drs. Guangzhong Wang, Zhen Shao and the anonymous reviewers ’ neighborhood and its known target gene GADD45A. Three of JUND s for helpful comments. neighborhood genes with strongest co-expression with GADD45 were chosen for display. (PDF 93 kb) Funding Additional file 2: Table S1. TFs deregulated in each lung cancer This work was supported by the National Basic Research Program of China dataset. Columns are: DS (dataset), DIR (direction of regulation), TF, lfc 2017YFA0505500 (H.J.), the Strategic Priority Research Program of the (log2 fold change), p (differential t test p value), ntargs (number of Chinese Academy of Sciences XDB19000000 (H.J.), the National Natural targets deregulated) and targs (targets deregulated). Table S2. GO terms Science Foundation of China 81602459 (Z.F.), 81430066 (H.J.), 91731314 (H.J.), enriched in targets of each TF. Columns are: DS (dataset), DIR (direction 31621003 (H.J.). of regulation), TF, GO, Term, Annotated (Number of genes annotated and recognized in GO term), GOI (Number of deregulated targets for each TF), Hits (TF targets annotated in the GO term), OR (odds ratio), pFisher Availability of data and materials (Fisher’s Exact Test p value) and Genes (Gene Hits). Table S3. Evaluation The TF regulatory network is available online (https://github.com/celreg/tf-lc). of SVM parameters. Different parameter combinations were used to set The datasets used and/or analyzed are available from public resources and/ up the OC-SVM model for training. Each model was used to predict a TF or the corresponding authors. network, which was then applied to the LUSC dataset to see if the two positive control TFs (TP63 and SOX2) can be recalled. Table S4. Evaluation Authors' contributions of differential analysis parameters. Several combinations of log2fc and q HJ and ZF designed the project. SZ and ZF performed the computational value thresholds were applied to determine up- and down-regulated genes analyses. SZ, LM, HJ and ZF interpreted the data and wrote the manuscript. in the LUSC dataset, which are further used for TF identification. The All authors read and approved the final manuscript. identified TFs are compared with each other to evaluate the procedural robustness upon the various parameter choices. Top left and right: Ethics approval and consent to participate summarizing tables; Bottom: detailed tables of TFs identified in each Not applicable. parameter combination. (XLS 3594 kb) Additional file 3: Figure S2. Down-regulation of TFs in LUSC. (A) Competing interests Consistency of down-regulated TFs identified in the LUSC and LUSC2 The authors declare that they have no competing interests. Zhang et al. BMC Genomics (2018) 19:435 Page 12 of 13

Publisher’sNote 22. Webb AE, Pollina EA, Vierbuchen T, Urbán N, Ucar D, Leeman DS, et al. Springer Nature remains neutral with regard to jurisdictional claims in FOXO3 Shares Common Targets with ASCL1 Genome-wide and Inhibits published maps and institutional affiliations. ASCL1-Dependent Neurogenesis. Cell Reports. 2013;4:477–91. 23. Machanick P, Bailey TL. MEME-ChIP: motif analysis of large DNA datasets. Author details Bioinformatics. 2011;27:1696–7. 1School of Life Sciences and Biotechnology, Shanghai Jiao Tong University, 24. Grant CE, Bailey TL, Noble WS. FIMO: scanning for occurrences of a given Shanghai 200240, People’s Republic of China. 2State Key Laboratory of Cell motif. Bioinformatics. 2011;27:1017–8. Biology, Shanghai, China. 3CAS Center for Excellence in Molecular Cell 25. Lonsdale J, Thomas J, Salvatore M, Phillips R, Lo E, Shad S, et al. The Science, Shanghai, China. 4Innovation Center for Cell Signaling Network, Genotype-Tissue Expression (GTEx) project. Nat Genet. 2013;45:580–5. Institute of Biochemistry and Cell Biology, Shanghai 200031, China. 5School 26. Law CW, Chen Y, Shi W, Smyth GK. voom: precision weights unlock linear of Life Science and Technology, Shanghai Tech University, Shanghai 200120, model analysis tools for RNA-seq read counts. Genome Biol. 2014;15:R29. 6 China. Shanghai Institutes for Biological Sciences, Chinese Academy of 27. Orchard S, Ammari M, Aranda B, Breuza L, Briganti L, Broackes-Carter F, et al. Science, Shanghai 200031, China. The MIntAct project—IntAct as a common curation platform for 11 molecular interaction databases. Nucleic Acids Res. 2014;42:D358–63. Received: 11 October 2017 Accepted: 25 May 2018 28. Schölkopf B, Williamson RC, Smola AJ, Shawe-Taylor J, Platt JC. Support vector method for novelty detection. Advances in neural information processing systems. 2000. p. 582–588. 29. Schölkopf B, Platt JC, Shawe-Taylor J, Smola AJ, Williamson RC. Estimating References the Support of a High-Dimensional Distribution. Neural Computation. 1. Green MR. Targeting Targeted Therapy. New England J Medicine. 2001;13:1443–71. 2004;350:2191–3. 30. Han H, Shim H, Shin D, Shim JE, Ko Y, Shin J, et al. TRRUST: a reference 2. Lynch TJ, Bell DW, Sordella R, Gurubhagavatula S, Okimoto RA, Brannigan database of human transcriptional regulatory interactions. Sci Rep. BW, et al. Activating Mutations in the Epidermal Growth Factor Receptor 2015;5:srep11432. Underlying Responsiveness of Non–Small-Cell Lung Cancer to Gefitinib. New England J Medicine. 2004;350:2129–39. 31. Zhao F, Xuan Z, Liu L, Zhang MQ. TRED: a Transcriptional Regulatory 3. Pao W, Girard N. New driver mutations in non-small-cell lung cancer. Element Database and a platform for in silico gene regulation studies. – Lancet Oncol. 2011;12:175–80. Nucleic Acids Res. 2005;33:D103 7. 4. Carlsson P, Mahlapuu M. Forkhead Transcription Factors: Key Players in 32. Montgomery SB, Griffith OL, Sleumer MC, Bergman CM, Bilenky M, Development and Metabolism. Dev Biol. 2002;250:1–23. Pleasance ED, et al. ORegAnno: an open access database and curation 5. Lee TI, Young RA. Transcriptional Regulation and Its Misregulation in system for literature-derived promoters, transcription factor binding sites – Disease. Cell. 2013;152:1237–51. and regulatory variation. Bioinformatics. 2006;22:637 40. 6. Voss TC, Hager GL. Dynamic regulation of transcriptional states by 33. Khan SS, Madden MG. A Survey of Recent Trends in One Class chromatin and transcription factors. Nat Rev Genet. 2014;15:69–81. Classification. Proceedings of the 20th Irish Conference on Artificial 7. Palazon A, Goldrath AW, Nizet V, Johnson RS. HIF Transcription Factors, Intelligence and Cognitive Science. Berlin, Heidelberg: Springer-Verlag; – Inflammation, and Immunity. Immunity. 2014;41:518–28. 2010. p. 188 197. 8. Almalki SG, Agrawal DK. Key transcription factors in the differentiation of 34. Gontan C, de Munck A, Vermeij M, Grosveld F, Tibboel D, Rottier R. Sox2 is mesenchymal stem cells. Differentiation. 2016;92:41–51. important for two crucial processes in lung development: Branching 9. Takahashi K, Yamanaka S. Induction of Pluripotent Stem Cells from morphogenesis and epithelial cell differentiation. Developmental Biology. – Mouse Embryonic and Adult Fibroblast Cultures by Defined Factors. 2008;317:296 309. Cell. 2006;126:663–76. 35. Hussenet T, Dali S, Exinger J, Monga B, Jost B, Dembelé D, et al. SOX2 Is an 10. Takahashi K, Tanabe K, Ohnuki M, Narita M, Ichisaka T, Tomoda K, et al. Oncogene Activated by Recurrent 3q26.3 Amplifications in Human Lung Induction of Pluripotent Stem Cells from Adult Human Fibroblasts by Squamous Cell Carcinomas. PLOS ONE. 2010;5:e8960. Defined Factors. Cell. 2007;131:861–72. 36. Ferone G, Song J-Y, Sutherland KD, Bhaskaran R, Monkhorst K, Lambooij 11. Yamanaka S. Strategies and new developments in the Generation of J-P, et al. SOX2 Is the Determining Oncogenic Switch in Promoting Patient-Specific Pluripotent Stem Cells. Cell Stem Cell. 2007;1:39–49. Lung Squamous Cell Carcinoma from Different Cells of Origin. Cancer 12. Weissman IL, Anderson DJ, Gage F. Stem and Progenitor Cells: Origins, Cell. 2016;30:519–32. Phenotypes, Lineage Commitments, and Transdifferentiations. Annu Rev 37. Kang JU, Koo SH, Kwon KC, Park JW. Frequent silence of chromosome 9p, Cell Dev Biol. 2001;17:387–403. homozygous DOCK8, DMRT1 and DMRT3 deletion at 9p24.3 in squamous 13. Lamouille S, Xu J, Derynck R. Molecular mechanisms of epithelial– cell carcinoma of the lung. Int J Oncol. 2010;37:327–35. mesenchymal transition. Nat Rev Mol Cell Biol. 2014;15:178–96. 38. Lo KC, Stein LC, Panzarella JA, Cowell JK, Hawthorn L. Identification of genes 14. Thiery JP. Epithelial–mesenchymal transitions in tumour progression. Nat involved in squamous cell carcinoma of the lung using synchronized data Rev Cancer. 2002;2:442–54. from DNA copy number and transcript expression profiling analysis. Lung 15. Bass AJ, Watanabe H, Mermel CH, Yu S, Perner S, Verhaak RG, et al. SOX2 is Cancer. 2008;59:315–31. an amplified lineage-survival oncogene in lung and esophageal squamous 39. Andersson LS, Larhammar M, Memic F, Wootz H, Schwochow D, Rubin C-J, cell carcinomas. Nat Genet. 2009;41:1238–42. et al. Mutations in DMRT3 affect locomotion in horses and spinal circuit 16. Massion PP, Taflan PM, Rahman SMJ, Yildiz P, Shyr Y, Edgerton ME, et al. function in mice. Nature. 2012;488:642–6. Significance of p63 Amplification and Overexpression in Lung Cancer 40. Santiago L, Daniels G, Wang D, Deng F-M, Lee P. Wnt signaling pathway Development and Prognosis. Cancer Res. 2003;63:7113–21. protein LEF1 in cancer, as a biomarker for prognosis and a target for 17. Crum CP, McKeon FD. p63 in Epithelial Survival, Germ Cell Surveillance, and treatment. Am J Cancer Res. 2017;7:1389–406. Neoplasia. Annu Rev Pathol. 2010;5:349–71. 41. Nguyen DX, Chiang AC, Zhang XH-F, Kim JY, Kris MG, Ladanyi M, et al. 18. Han X, Li F, Fang Z, Gao Y, Li F, Fang R, et al. Transdifferentiation of lung WNT/TCF Signaling through LEF1 and HOXB9 Mediates Lung adenocarcinoma in mice with Lkb1 deficiency to squamous cell carcinoma. Adenocarcinoma Metastasis. Cell. 2009;138:51–62. Nat Commun. 2014;5:3261–1. 42. Mathas S, Janz M, Hummel F, Hummel M, Wollert-Wulf B, Lusatis S, et al. 19. George J, Lim JS, Jang SJ, Cun Y, Ozretić L, Kong G, et al. [SCLC]Comprehensive Intrinsic inhibition of transcription factor E2A by HLH proteins ABF-1 and genomic profiles of small cell lung cancer. Nature. 2015;524:47–53. Id2 mediates reprogramming of neoplastic B cells in Hodgkin lymphoma. 20. Mathelier A, Fornes O, Arenillas DJ, Chen C, Denay G, Lee J, et al. JASPAR Nat Immunol. 2006;7:207–15. 2016: a major expansion and update of the open-access database of 43. Weir BA, Woo MS, Getz G, Perner S, Ding L, Beroukhim R, et al. transcription factor binding profiles. Nucleic Acids Res. 2016;44:D110–5. Characterizing the cancer genome in lung adenocarcinoma. Nature. 21. Watanabe H, Francis JM, Woo MS, Etemad B, Lin W, Fries DF, et al. 2007;450:893. Integrated cistromic and expression analysis of amplified NKX2-1 in lung 44. Li CM-C, Gocheva V, Oudin MJ, Bhutkar A, Wang SY, Date SR, et al. Foxa2 adenocarcinoma identifies LMO3 as a functional transcriptional target. and Cdx2 cooperate with Nkx2-1 to inhibit lung adenocarcinoma Genes & Development. 2013;27:197–210. metastasis. Genes Dev. 2015;29:1850–62. Zhang et al. BMC Genomics (2018) 19:435 Page 13 of 13

45. Mu D. The Complexity of Thyroid Transcription Factor 1 with Both Pro- and Anti-oncogenic Activities. J Biol Chem. 2013;288:24992–5000. 46. Yamaguchi T, Hosono Y, Yanagisawa K, Takahashi T. NKX2-1/TTF-1: An Enigmatic Oncogene that Functions as a Double-Edged Sword for Cancer Cell Survival and Progression. Cancer Cell. 2013;23:718–23. 47. Osada H, Tatematsu Y, Yatabe Y, Horio Y, Takahashi T. ASH1 Gene Is a Specific Therapeutic Target for Lung Cancers with Neuroendocrine Features. Cancer Res. 2005;65:10680–5. 48. Jiang T, Collins BJ, Jin N, Watkins DN, Brock MV, Matsui W, et al. Achaete- Scute Complex Homologue 1 Regulates Tumor-Initiating Capacity in Human Small Cell Lung Cancer. Cancer Res. 2009;69:845–54. 49. Augustyn A, Borromeo M, Wang T, Fujimoto J, Shao C, Dospoy PD, et al. ASCL1 is a lineage oncogene providing therapeutic targets for high-grade neuroendocrine lung cancers. PNAS. 2014;111:14788–93. 50. Otero K, Vecchi A, Hirsch E, Kearley J, Vermi W, Prete AD, et al. Nonredundant role of CCRL2 in lung dendritic cell trafficking. Blood. 2010;116:2942–9. 51. Hann CL, Rudin CM. Fast, hungry and unstable: finding the Achilles’ heel of small-cell lung cancer. Trends Mol Med. 2007;13:150–7. 52. Jiang B, Zhang MQ, Zhang XOSCAR. One-class SVM for accurate recognition of cis-elements. Bioinformatics. 2007;23:2823–8. 53. Mathelier A, Wasserman WW. The Next Generation of Transcription Factor Binding Site Prediction. PLOS Comp Biol. 2013;9:e1003214. 54. Ghandi M, Lee D, Mohammad-Noori M, Beer MA. Enhanced Regulatory Sequence Prediction Using Gapped k-mer Features. PLOS Comput Biol. 2014;10:e1003711. 55. Alipanahi B, Delong A, Weirauch MT, Frey BJ. Predicting the sequence specificities of DNA- and RNA-binding proteins by deep learning. Nat Biotech. 2015;33:831–8. 56. Zhou J, Troyanskaya OG. Predicting effects of noncoding variants with deep learning–based sequence model. Nat Methods. 2015;12:931–4. 57. Keilwagen J, Grau J. Varying levels of complexity in transcription factor binding motifs. Nucleic Acids Res. 2015;43:e119–9. 58. Yang J, Ramsey SA. A DNA shape-based regulatory score improves position- weight matrix-based recognition of transcription factor binding sites. Bioinformatics. 2015;31:3445–50. 59. Qin Q, Feng J. Imputation for transcription factor binding predictions based on deep learning. PLOS Comput Biol. 2017;13:e1005403. 60. Siggers T, Gordân R. Protein–DNA binding: complexities and multi-protein codes. Nucleic Acids Res. 2014;42:2099–111. 61. Slattery M, Zhou T, Yang L, Dantas Machado AC, Gordân R, Rohs R. Absence of a simple code: how transcription factors read the genome. Trends Biochem Sci. 2014;39:381–99. 62. Qian J, Lin J, Luscombe NM, Yu H, Gerstein M. Prediction of regulatory networks: genome-wide identification of transcription factor targets from gene expression data. Bioinformatics. 2003;19:1917–26. 63. Holloway DT, Kon M, DeLisi C. Machine learning for regulatory analysis and transcription factor target prediction in yeast. Syst Synth Biol. 2007;1:25–46.